Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A method for improving efficiency of an image search, the method comprising: receiving, by a first processor, one or more images; sending, by the first processor, a query to a plurality of memory modules, wherein the query is for a search of memory for a matching image to the one or more images; performing, by a processor-in-memory (PIM)of each respective memory module from the plurality of memory modules, a search of images stored in a memory of the respective memory module, in response to the query received; sending, by the PIM of each respective memory model, the matching image obtained from the search to the first processor; and performing, by the first processor, a comparison of the matching image received from each of the plurality of memory modules to the one or more images, wherein the PIM of each respective memory module is configured to perform the search of images by: processing a global descriptor of the one or more images to obtain one or more similarity measurements; and comparing local descriptors of one or more images whose similarity measurements are above a threshold to the images stored in the memory of the respective memory module to obtain the matching images, wherein the images stored in each respective memory module of the plurality of memory modules are different.
Image search efficiency improvement. This invention addresses the challenge of efficiently searching large image datasets. The method involves a system with a first processor and multiple memory modules, each containing its own processor-in-memory (PIM). The first processor receives one or more input images. It then sends a query to all the memory modules, requesting them to search their stored images for matches. Each PIM within a memory module independently searches its local memory for matching images. To perform this search, each PIM first processes a global descriptor of the input images to calculate similarity measurements. Images with similarity measurements exceeding a predefined threshold are then further analyzed. The PIM compares local descriptors of these promising images against the images stored in its memory to identify matching images. Each PIM sends any matching images it finds back to the first processor. Finally, the first processor compares the matching images received from all the memory modules to the original input images to determine the overall best matches. Crucially, the images stored in each memory module are distinct, ensuring a distributed and comprehensive search.
2. The method of claim 1 , further comprising uploading and classifying the images stored in the memory of the at least one memory module.
This invention relates to image processing systems that classify images stored in memory. The system includes at least one memory module storing images and a processing unit configured to analyze these images. The processing unit performs image classification by extracting features from the images and categorizing them based on predefined criteria. The classification process may involve machine learning algorithms or rule-based systems to determine the content or attributes of each image. The system may also include a user interface for managing the classification process, allowing users to review, edit, or refine the classifications. Additionally, the system may support uploading new images to the memory module, where they are automatically processed and classified. The classification results can be stored alongside the images for future reference or retrieval. This invention addresses the need for automated image organization, particularly in large datasets where manual classification is impractical. The system improves efficiency by reducing the time and effort required to categorize images, making it useful in applications such as digital asset management, medical imaging, and surveillance. The classification process may also include error handling to ensure accuracy, such as flagging ambiguous or low-confidence classifications for manual review.
3. The method of claim 2 wherein the images are stored on the plurality of memory modules based on the classifying.
This invention relates to a system for managing and storing digital images based on their content. The problem addressed is the inefficient storage and retrieval of images in large datasets, where images are often stored without organization, making it difficult to locate specific images later. The invention provides a solution by classifying images into categories and distributing them across multiple memory modules according to their classifications. The method involves analyzing images to extract features or metadata that define their content, such as objects, scenes, or other attributes. These features are then used to classify the images into predefined categories. Once classified, the images are stored on different memory modules in a distributed storage system, with each module designated for a specific category. This ensures that images are stored in an organized manner, improving retrieval efficiency. The system may also include a preprocessing step to enhance image quality or extract additional metadata before classification. The classification process may use machine learning models or rule-based systems to determine the appropriate category for each image. The storage distribution ensures that frequently accessed categories are stored on high-performance memory modules, while less frequently accessed categories are stored on lower-performance modules, optimizing storage resources. By storing images based on their classifications, the system enables faster retrieval of relevant images and reduces the time required to search through large datasets. This approach is particularly useful in applications such as digital libraries, medical imaging, and surveillance systems where efficient image management is critical.
4. The method of claim 2 wherein the images are stored on a same memory module when the images have a same classification.
This invention relates to image storage and classification systems, specifically addressing the challenge of efficiently organizing and retrieving images based on their content or metadata. The method involves classifying images into categories and storing them on the same memory module when they share the same classification. This approach optimizes storage access by grouping related images together, reducing search time and improving retrieval efficiency. The classification can be based on various criteria, such as content, metadata, or user-defined tags. By storing similarly classified images in the same memory module, the system minimizes the need to access multiple storage locations, enhancing performance in applications like image databases, digital libraries, or content management systems. The method ensures that frequently accessed or related images are co-located, reducing latency and improving overall system responsiveness. This technique is particularly useful in large-scale image storage systems where quick retrieval and organization are critical.
5. The method of claim 1 wherein the search of images performed by the at least one memory module includes performing a global descriptor computation, a local descriptor computation, and a verification determination.
This invention relates to image search and retrieval systems, specifically improving the efficiency and accuracy of searching large image databases. The problem addressed is the computational complexity and inefficiency of traditional image search methods, which often rely on either global or local descriptors alone, leading to either high computational cost or reduced accuracy. The method involves a multi-stage search process within a memory module to enhance image retrieval performance. First, a global descriptor computation is performed to quickly narrow down potential matches by analyzing high-level features of the images. This step reduces the search space by identifying images that broadly resemble the query. Next, a local descriptor computation is applied to the narrowed set of images, focusing on finer details and specific regions to refine the search results. Finally, a verification determination step is used to confirm the most accurate matches by cross-checking the global and local descriptors, ensuring high precision in the final results. By combining global and local descriptors with a verification step, the method balances computational efficiency with search accuracy, making it suitable for large-scale image databases where both speed and precision are critical. This approach is particularly useful in applications such as content-based image retrieval, visual search engines, and automated image categorization.
6. The method of claim 1 wherein the comparison of the comparison includes ranking the matching images received from the at least one memory module.
This invention relates to image processing and retrieval systems, specifically improving the efficiency and accuracy of matching and ranking images from a database. The problem addressed is the challenge of efficiently retrieving and ranking relevant images from large datasets based on similarity or other criteria, which is computationally intensive and often results in suboptimal performance. The method involves comparing a query image against a collection of images stored in at least one memory module. The comparison process includes analyzing the images to identify matches based on predefined criteria, such as visual similarity, metadata, or other features. The key innovation is the inclusion of a ranking step, where the matching images are sorted or prioritized according to their relevance or similarity to the query image. This ranking may be based on factors such as confidence scores, similarity metrics, or user-defined preferences. The ranked results are then provided to the user or another system for further processing or display. The ranking step ensures that the most relevant images are presented first, improving user experience and reducing the need for manual filtering. This method is particularly useful in applications such as image search engines, content-based retrieval systems, or automated image analysis tools. The system may also include additional features, such as dynamic adjustment of ranking criteria based on user feedback or real-time performance metrics. The overall goal is to enhance the efficiency and accuracy of image retrieval in large-scale databases.
7. The method of claim 6 , further comprising constructing, by the first processor, a final rank of the matching images received from multiple memory modules.
The invention relates to a system for ranking and retrieving matching images from multiple memory modules in a distributed computing environment. The problem addressed is efficiently organizing and prioritizing image search results when data is stored across multiple memory modules, ensuring fast and accurate retrieval of relevant images. The method involves a first processor receiving matching images from multiple memory modules, where each module may have processed a subset of a query to identify relevant images. The processor then constructs a final rank of these matching images based on their relevance or other criteria. This ranking step ensures that the most relevant images are prioritized in the final output, improving user experience in applications such as image search, recommendation systems, or content filtering. The system may include multiple processors, where a second processor distributes the query to the memory modules and a third processor may further refine the ranking based on additional factors like user preferences or contextual data. The memory modules store image data and perform initial matching operations, while the processors handle coordination, distribution, and final ranking tasks. This distributed approach enhances scalability and performance in large-scale image retrieval systems.
8. A processor for improving efficiency of an image search, the processor comprising: circuitry configured to send a query for a matching image to a plurality of memory modules, wherein the query includes one or more search images to be matched; circuitry configured to receive search results from the plurality of memory modules, wherein the search results are received in response to the query; and circuitry configured to perform a comparison of the search results received from the plurality of memory modules, wherein the search results are generated by a processor in memory (PIM) of each of the plurality of memory modules performing a search of a respective memory module, wherein the search includes: processing a global descriptor of the one or more search images to obtain one or more similarity measurements; and comparing local descriptors of one or more search images whose similarity measurements are above a threshold to images stored in the memory of the respective memory module to obtain the matching image, wherein the images stored in each respective memory module of the plurality of memory modules are different.
This invention relates to improving the efficiency of image search operations using processor-in-memory (PIM) technology. The problem addressed is the computational overhead and latency associated with traditional image search methods, which often require transferring large image datasets between memory and processing units. The solution involves a processor designed to distribute image search tasks across multiple memory modules, each equipped with PIM capabilities. The processor sends a query containing one or more search images to these memory modules, which independently perform local searches. Each memory module processes a global descriptor of the search images to generate similarity measurements. If these measurements exceed a predefined threshold, the module further compares local descriptors of the search images against images stored in its own memory to identify matches. The images stored in each memory module are distinct, allowing parallel processing and reducing the overall search time. This approach leverages the parallelism of PIM to accelerate image retrieval while minimizing data transfer between memory and processing units. The system is particularly useful in applications requiring fast and efficient image matching, such as large-scale image databases or real-time search systems.
9. The processor of claim 8 , further comprising circuitry configured to upload and classify the images stored in the memory of the plurality of memory modules.
This invention relates to a processor system for managing and processing images stored in a distributed memory architecture. The system addresses the challenge of efficiently handling large volumes of image data across multiple memory modules, ensuring fast access, classification, and upload capabilities. The processor includes circuitry designed to upload images from a plurality of memory modules and classify them based on predefined criteria. The memory modules are interconnected in a distributed network, allowing parallel processing and reducing bottlenecks. The classification circuitry categorizes images into groups, enabling optimized storage, retrieval, and analysis. This system is particularly useful in applications requiring real-time image processing, such as surveillance, medical imaging, or autonomous systems, where rapid data handling and organization are critical. The distributed memory architecture enhances scalability and fault tolerance, ensuring reliable performance even with high data loads. The processor's ability to classify images on-the-fly improves efficiency in applications where immediate categorization is necessary, such as content filtering or object recognition. By integrating upload and classification functions within the processor, the system minimizes latency and maximizes throughput, making it suitable for high-performance computing environments.
10. The processor of claim 8 , further comprising circuitry configured to rank the search results received from the plurality of memory modules.
A system for optimizing search operations in a distributed memory architecture addresses the challenge of efficiently retrieving and ranking data from multiple memory modules. The system includes a processor with circuitry designed to distribute search queries across a plurality of memory modules, each storing a portion of a dataset. The processor further includes circuitry to receive search results from the memory modules and rank those results based on relevance or other criteria. The ranking circuitry may use algorithms such as relevance scoring, proximity matching, or other ranking techniques to prioritize the results. The system ensures efficient data retrieval by parallelizing search operations across distributed memory modules, reducing latency and improving performance in large-scale data processing environments. The ranking step consolidates results from different memory modules into a unified, ordered output, enhancing usability for applications requiring prioritized data retrieval. This approach is particularly useful in systems where data is distributed across multiple memory units, such as in distributed databases, cloud storage, or high-performance computing environments. The system optimizes both search efficiency and result quality by leveraging parallel processing and intelligent ranking mechanisms.
11. The processor of claim 10 , further comprising circuitry configured to construct a final rank of the search results.
A system for ranking search results involves a processor with circuitry that processes search queries and generates ranked search results. The processor includes circuitry to receive a search query, retrieve relevant documents or data from a database, and generate an initial ranking of the search results based on relevance metrics. Additional circuitry is configured to refine the ranking by applying user-specific preferences, historical data, or contextual factors. The system further includes circuitry to construct a final rank of the search results, which may involve combining multiple ranking criteria, applying machine learning models, or optimizing for user engagement. The final ranking is then output to the user, ensuring the most relevant and personalized results are displayed. This approach improves search accuracy and user satisfaction by dynamically adjusting rankings based on various factors.
12. A system for improving efficiency of an image search, the system comprising: a plurality of memory modules, where each of the plurality of memory modules includes a memory and a processor-in-memory (PIM); a processor that is communicatively coupled to the plurality of memory modules, wherein the processor comprises: circuitry configured to receive one or more search images; circuitry configured to send a query to the plurality of memory modules, wherein the query includes the one or more search images; circuitry configured to receive search results from the plurality of memory modules; and circuitry configured to perform a comparison of the search results received from the plurality of memory modules; and wherein the PIM of each respective memory module is configured to perform the search by: processing a global descriptor of the one or more search images to obtain one or more similarity measurements; and comparing local descriptors of one or more search images whose similarity measurements are above a threshold to images stored in the respective memory module to obtain the search results, wherein the images stored in each respective memory module of the plurality of memory modules are different.
The system improves the efficiency of image search by leveraging processor-in-memory (PIM) technology to offload computational tasks from a central processor to memory modules. Traditional image search systems often suffer from bottlenecks due to the high computational cost of comparing large image datasets, particularly when using both global and local descriptors for accurate matching. This system addresses the problem by distributing the search workload across multiple memory modules, each equipped with a PIM and storing a distinct subset of images. The central processor receives one or more search images and sends a query containing these images to the memory modules. Each PIM processes the global descriptors of the search images to compute similarity measurements, filtering out images with low similarity. Only images with similarity measurements above a predefined threshold undergo further comparison using local descriptors, reducing unnecessary computations. The memory modules return search results to the central processor, which then compares and consolidates the results. By distributing the search process and leveraging PIM for localized computations, the system enhances search efficiency and reduces latency. The images stored in each memory module are unique, ensuring a balanced workload distribution.
13. The system of claim 12 , wherein the processor further comprises circuitry configured to upload and classify the images stored in the memory of the plurality of memory modules.
This invention relates to a distributed image processing system designed to efficiently manage and analyze large volumes of image data across multiple memory modules. The system addresses the challenge of handling high-throughput image data in environments where centralized processing is impractical or inefficient, such as in large-scale surveillance, medical imaging, or autonomous vehicle applications. The system includes a processor with specialized circuitry that coordinates the storage, retrieval, and processing of images across a network of distributed memory modules. Each memory module stores a subset of the total image dataset, and the processor dynamically allocates tasks to optimize performance. The circuitry is configured to upload images into the memory modules and classify the stored images based on predefined criteria, such as content, metadata, or user-defined tags. This classification enables faster retrieval and targeted analysis, reducing latency and improving system responsiveness. The processor may also include additional circuitry for real-time image analysis, such as object detection, pattern recognition, or anomaly identification. The distributed architecture ensures scalability, allowing the system to expand by adding more memory modules without significant performance degradation. The classification functionality enhances data organization, making it easier to search and retrieve specific images for further processing or user access. This system is particularly useful in applications requiring high-speed, distributed image processing with minimal central coordination.
14. The system of claim 13 wherein the images are stored on the plurality of memory modules based on a classification of the images stored in the memory of the plurality of memory modules.
The invention relates to a distributed image storage system that organizes and retrieves images based on their classification. The system addresses the challenge of efficiently managing large volumes of images in a distributed memory architecture, ensuring fast access and optimal storage utilization. The system includes multiple memory modules, each capable of storing images and performing classification tasks. When an image is received, it is classified into one or more categories, and the classification determines which memory module will store the image. This classification-based distribution ensures that related images are stored together, improving retrieval efficiency and reducing search times. The system dynamically adjusts storage allocation based on the classification of images already present in the memory modules, balancing load and optimizing performance. Additionally, the system may include mechanisms to handle image updates, deletions, and reclassification, ensuring data consistency across the distributed storage. The classification process may involve machine learning models or predefined rules to categorize images based on content, metadata, or other attributes. By storing images in a structured manner according to their classification, the system enhances scalability and accessibility in large-scale image storage applications.
15. The system of claim 13 wherein the images stored on a same memory module when the images have a same classification .
A system for organizing digital images based on classification involves storing images on a memory module according to their classification. The system includes a memory module configured to store digital images and a classification module that assigns a classification to each image. The classification module processes image data to determine attributes such as content, metadata, or user-defined tags, then categorizes the images accordingly. The system ensures that images with the same classification are stored on the same memory module, optimizing storage efficiency and retrieval speed. This approach reduces fragmentation and improves access times by grouping related images together. The system may also include a user interface for managing classifications and viewing organized images. The classification module can dynamically update classifications based on new data or user input, ensuring the system remains accurate and adaptable. This method enhances digital image management by automating organization and maintaining logical storage structures.
16. The system of claim 12 wherein the search of images performed by the at least one memory module includes performing a global descriptor computation, a local descriptor computation, and a verification determination.
The invention relates to an image search system designed to efficiently retrieve relevant images from a database based on query inputs. The system addresses the challenge of accurately and quickly identifying images that match or are similar to a given query image, which is particularly useful in applications like visual search, content-based image retrieval, and automated image organization. The system includes at least one memory module that performs a multi-stage search process. First, it computes a global descriptor for each image, which captures high-level, holistic features of the entire image. This step helps narrow down the search to a subset of images that are broadly similar to the query. Next, the system computes local descriptors, which focus on specific regions or keypoints within the images, providing finer-grained comparisons. Finally, a verification determination step is performed to confirm the relevance of the candidate images, ensuring high accuracy in the search results. This multi-stage approach balances computational efficiency with precision, making the system suitable for large-scale image databases. The system may also include additional components, such as a query input module and a display module, to facilitate user interaction and result presentation.
17. The system of claim 12 , wherein the processor further comprises circuitry configured to rank the search results received from the plurality of memory modules.
The system is designed for efficient data retrieval in distributed memory architectures, addressing challenges in managing and accessing large datasets across multiple memory modules. The system includes a processor with circuitry that coordinates data searches across a plurality of memory modules, each storing portions of a dataset. The processor sends search queries to these modules, which return relevant data segments. The circuitry in the processor then ranks the search results based on predefined criteria, such as relevance, recency, or other metadata attributes, to prioritize the most pertinent results. This ranking mechanism ensures that the most valuable data is presented first, improving efficiency in data retrieval tasks. The system may also include additional features, such as parallel processing of queries, load balancing across memory modules, and adaptive ranking algorithms that adjust based on usage patterns or user preferences. The overall goal is to optimize search performance in distributed memory environments, reducing latency and enhancing accuracy in data retrieval operations.
18. The system of claim 17 , wherein the processor further comprises circuitry configured to construct a final rank of the search results received from the plurality of memory modules.
The system relates to search result ranking in distributed memory architectures, addressing the challenge of efficiently combining and ranking search results from multiple memory modules to improve accuracy and relevance. The system includes a processor with circuitry that constructs a final rank of search results received from a plurality of memory modules. The processor circuitry is configured to process search queries and distribute them across the memory modules, which independently generate and return search results. The system further includes a ranking mechanism that evaluates the relevance of each result based on predefined criteria, such as relevance scores, recency, or user preferences. The processor circuitry then aggregates and normalizes these results, applying a ranking algorithm to produce a final ranked list. This ensures that the most relevant results are prioritized, even when derived from distributed memory sources. The system may also incorporate machine learning models to refine ranking over time based on user interactions. By dynamically adjusting the ranking process, the system enhances search performance in large-scale distributed systems, improving user experience and efficiency.
19. A method implemented for improving efficiency of an image search that is in a processor-in-memory (PIM) system having a first processor and a plurality of memory modules, the method comprising: uploading and classifying one or more images by their image data; partitioning and storing the images on the plurality of memory modules based upon the classification of the image data; receiving one or more query images by the first processor; sending, by the first processor, a query to the plurality of memory modules, wherein the query includes the one or more query images; performing, by a PIM of each respective memory module of the plurality of memory modules, a search of images stored in a memory of the respective memory module, wherein the search of images is performed in response to the query; sending, by each memory module of the plurality of memory modules, search results to the first processor; and performing, by the first processor, a comparison of the search results from the plurality of memory modules, wherein the search of images comprises: processing a global descriptor of the one or more query images to obtain one or more similarity measurements; and comparing local descriptors of one or more query images whose similarity measurements are above a threshold to images stored in the respective memory module to obtain the search results.
The invention relates to improving the efficiency of image search in a processor-in-memory (PIM) system. Traditional image search systems often suffer from high latency and computational overhead when processing large datasets, as queries must be sent to a central processor for comparison against stored images. This invention addresses these inefficiencies by leveraging distributed processing within memory modules to accelerate image retrieval. The system includes a first processor and multiple memory modules, each equipped with PIM capabilities. Images are uploaded and classified based on their data, then partitioned and stored across the memory modules according to their classifications. When a query image is received, the first processor sends the query to all memory modules. Each memory module independently searches its stored images using its PIM capabilities. The search process involves two stages: first, global descriptors of the query images are processed to generate similarity measurements. Only query images with similarity measurements above a predefined threshold proceed to the second stage, where their local descriptors are compared against images in the respective memory module. The results from each module are then sent back to the first processor, which consolidates and compares them to produce the final search output. This distributed approach reduces latency and computational load on the central processor, improving overall search efficiency.
20. The method of claim 19 wherein the images are stored on a same memory module when the images have a same classification.
A system and method for organizing and storing digital images based on their classification to improve retrieval efficiency. The method involves analyzing images to determine their classification, such as subject matter, content type, or metadata tags, and then storing images with the same classification on the same memory module. This reduces access time and improves system performance by grouping related images together, minimizing the need to search across multiple storage locations. The system may include a classification engine that processes images to extract relevant features or metadata, a storage controller that manages the distribution of images to appropriate memory modules, and a retrieval mechanism that quickly locates images by referencing their stored classification. The method ensures that frequently accessed or related images are stored in close proximity, optimizing storage and retrieval operations. This approach is particularly useful in large-scale image databases, such as those used in medical imaging, surveillance, or digital asset management, where efficient organization and rapid access are critical. The system may also include error handling to manage misclassification or storage failures, ensuring data integrity and reliability.
21. The method of claim 19 wherein the images from a same class are distributed among the plurality of memory modules.
This invention relates to distributed memory systems for image classification tasks, particularly in machine learning or computer vision applications. The problem addressed is the inefficiency in memory access when processing large datasets of classified images, where images of the same class are often stored together in a single memory module. This can lead to bottlenecks, as multiple processing units may need to access the same module simultaneously, causing delays and reducing system performance. The invention improves upon this by distributing images from the same class across multiple memory modules. This ensures that access requests for images of the same class are spread across different memory modules, reducing contention and improving parallel processing efficiency. The method involves assigning images to memory modules in a way that avoids clustering images of the same class in a single module, thereby balancing the load and optimizing memory bandwidth utilization. This approach is particularly useful in high-performance computing environments where large-scale image datasets are processed in parallel. The distribution strategy may involve hashing techniques, round-robin assignment, or other load-balancing algorithms to ensure even distribution. The result is faster retrieval and processing of images, leading to improved overall system throughput.
22. The method of claim 21 wherein the images from the same class are distributed among adjacent memory modules.
This invention relates to distributed memory systems for storing and retrieving image data, particularly in applications where images are categorized into classes. The problem addressed is the efficient distribution of image data across multiple memory modules to optimize access speed and reduce latency, especially when images belong to the same class and are frequently accessed together. The method involves organizing image data such that images from the same class are stored in adjacent memory modules. This ensures that when an image from a particular class is accessed, related images in the same class are also readily available in nearby memory locations, minimizing the need for long-distance data transfers within the memory system. The approach leverages spatial locality, where frequently co-accessed data is stored close together, improving performance in applications like image recognition, machine learning, or real-time processing systems. The method may also include techniques for dynamically adjusting the distribution of images based on access patterns, ensuring that frequently accessed classes remain optimally placed. Additionally, it may incorporate error correction or redundancy mechanisms to maintain data integrity while distributing images across multiple modules. The system can be implemented in hardware, software, or a combination of both, depending on the specific requirements of the application. The overall goal is to enhance memory access efficiency by reducing latency and improving throughput for image data retrieval operations.
23. The method of claim 19 wherein the search of images performed by the plurality of memory modules includes performing a global descriptor computation, a local descriptor computation, and a verification determination.
This invention relates to image search systems, specifically improving the efficiency and accuracy of searching large image databases. The problem addressed is the computational complexity and latency in retrieving relevant images from vast datasets, particularly when using distributed memory architectures. The method involves a distributed search system where multiple memory modules collaboratively process image queries. Each module performs a multi-stage analysis: first, a global descriptor computation extracts high-level features representing the entire image. Next, a local descriptor computation identifies detailed features from specific regions. Finally, a verification determination compares these descriptors to confirm matches. This hierarchical approach reduces redundant computations and improves search precision. The system optimizes performance by parallelizing these computations across memory modules, ensuring fast retrieval without sacrificing accuracy. The global descriptor provides a coarse filter to quickly narrow down candidates, while the local descriptor and verification refine results. This method is particularly useful in applications requiring real-time image retrieval, such as visual search engines or augmented reality systems. The distributed architecture scales efficiently with increasing data volumes, maintaining low latency even as the database grows.
24. The method of claim 23 wherein the computing of the global descriptor includes comparing the one or more query images to the images stored in the plurality of memory modules categorized as being in a same class as the one or more query images.
The invention relates to image retrieval systems, specifically improving the efficiency of computing global descriptors for query images by leveraging categorized image data. The problem addressed is the computational overhead in comparing query images against large datasets, particularly when the dataset is distributed across multiple memory modules. The solution involves computing a global descriptor for a query image by selectively comparing it only to images stored in memory modules categorized under the same class as the query image. This reduces the number of comparisons needed, as the system avoids processing irrelevant images from different classes. The method ensures that the global descriptor accurately represents the query image by focusing comparisons within the relevant subset of the dataset, thereby optimizing both speed and resource usage. The system may also involve pre-processing steps to categorize images into classes and distribute them across memory modules for efficient retrieval. This approach is particularly useful in large-scale image databases where fast and accurate retrieval is critical.
25. The method of claim 24 , further comprising computing the similarity measurement and comparing the similarity measurement to a threshold.
A system and method for analyzing data involves processing input data to generate a similarity measurement between two or more data sets. The method includes extracting features from the input data, where these features may be numerical, categorical, or derived from other data representations. The extracted features are then used to compute a similarity measurement, which quantifies how closely related the data sets are. This measurement can be based on various similarity metrics, such as Euclidean distance, cosine similarity, or other statistical or machine learning-based approaches. The computed similarity measurement is then compared to a predefined threshold to determine whether the data sets meet a specified level of similarity. If the similarity exceeds the threshold, the system may trigger further actions, such as flagging the data for review, merging the data sets, or generating an alert. The method may also include adjusting the threshold dynamically based on contextual factors or historical data to improve accuracy. This approach is useful in applications like fraud detection, recommendation systems, or data deduplication, where identifying similar data sets is critical. The system may operate in real-time or batch processing modes, depending on the application requirements.
26. The method of claim 25 wherein if the similarity measurement exceeds the threshold, the local descriptor is computed.
Technical Summary: This invention relates to image processing and computer vision, specifically to methods for efficiently computing local descriptors in images. The problem addressed is the computational cost of descriptor calculation, which can be resource-intensive when applied to every region of an image. The solution involves a conditional approach where local descriptors are only computed for regions where a similarity measurement exceeds a predefined threshold, thereby optimizing processing time and resources. The method first involves analyzing an image to identify regions of interest. For each region, a similarity measurement is computed, which quantifies how similar the region is to a reference or previously processed region. If this similarity measurement exceeds a threshold value, indicating sufficient distinctiveness or relevance, a local descriptor is then computed for that region. The descriptor is a compact representation of the region's visual features, useful for tasks like object recognition or image matching. If the similarity measurement does not exceed the threshold, the descriptor is not computed, skipping unnecessary computations. This approach reduces the overall computational load by avoiding descriptor calculations for regions that are too similar to others, thereby improving efficiency without sacrificing accuracy for relevant regions. The threshold can be adjusted based on application requirements, balancing between computational efficiency and descriptor coverage.
27. A non-transitory computer-readable medium having instructions recorded thereon that, when executed by a computing device, cause the computing device to perform operations to improve efficiency of an image search, the operations comprising: receiving one or more query images; sending a query to a plurality of memory modules, wherein the query includes the one or more query images; performing, by a processor in memory (PIM) of each of the plurality memory modules, a search of images stored in a respective memory of a respective memory module, wherein the search is performed in response to the query; performing a comparison of search received results received from the plurality of memory modules, wherein the search of images is performed by each PIM of the plurality of memory modules: processing a global descriptor of the one or more query images to obtain one or more similarity measurements; and comparing local descriptors of one or more query images whose similarity measurements are above a threshold to images stored in the memory of the respective memory module to obtain the search results.
This invention relates to improving the efficiency of image search systems, particularly by leveraging in-memory processing to reduce computational overhead. The problem addressed is the high latency and resource consumption in traditional image search systems, which often rely on centralized processing to compare query images against large datasets stored in external memory. The solution involves distributing the search workload across multiple memory modules, each equipped with a processor-in-memory (PIM) unit. When a query image is received, it is broadcast to all memory modules, which independently perform local searches using their PIM units. Each PIM processes a global descriptor of the query image to generate similarity measurements, filtering out low-relevance candidates. Only images with similarity measurements above a predefined threshold undergo a more detailed comparison using local descriptors, which are stored within the respective memory module. This approach minimizes data movement between memory and processor, reducing latency and energy consumption. The system efficiently scales with additional memory modules, as each operates in parallel to contribute to the final search results. The invention is particularly useful in applications requiring real-time image retrieval, such as visual search engines or augmented reality systems.
28. The non-transitory computer-readable medium of claim 27 , further comprising uploading and classifying the images stored in the plurality memory modules.
This invention relates to a system for managing and processing images stored in multiple memory modules. The system addresses the challenge of efficiently organizing and retrieving large volumes of image data distributed across different storage locations. The invention includes a method for uploading images to a plurality of memory modules, where each module may be located in different physical or logical storage devices. The system further classifies the uploaded images based on predefined criteria, such as content, metadata, or user-defined tags, to facilitate organized storage and retrieval. The classification process may involve machine learning algorithms or rule-based systems to automatically categorize images into relevant groups. The system ensures that images are properly indexed and accessible, improving searchability and usability. Additionally, the invention may include features for synchronizing image data across the memory modules to maintain consistency and integrity. The overall solution enhances the efficiency of image management in distributed storage environments, particularly in applications requiring large-scale image processing, such as cloud storage, digital asset management, or medical imaging systems.
29. The non-transitory computer-readable medium of claim 28 wherein the images from a same class are stored on the plurality of memory modules.
This invention relates to a data storage system for organizing and retrieving images based on their classification. The system addresses the challenge of efficiently storing and accessing large datasets of images, particularly when images belong to distinct categories or classes. The invention involves a non-transitory computer-readable medium containing instructions that, when executed, manage the storage of images across multiple memory modules. Specifically, the system ensures that images from the same class are stored together on the same memory module. This approach improves retrieval efficiency by reducing the need to search across multiple storage locations when accessing images of a particular class. The system may also include additional features, such as indexing or metadata management, to further enhance performance. By grouping related images on the same storage module, the invention optimizes both storage organization and retrieval speed, making it particularly useful in applications like image databases, machine learning datasets, or content management systems. The solution leverages distributed storage techniques to balance load and improve scalability while maintaining fast access to class-specific image data.
30. The non-transitory computer-readable medium of claim 28 wherein the images from a same class are stored on a same memory module.
This invention relates to a system for organizing and storing digital images in a memory system, specifically addressing the challenge of efficiently managing large datasets of images by categorizing and storing them in a structured manner. The system involves a method for classifying images into distinct classes based on their content or metadata, and then storing images belonging to the same class on the same memory module within a memory system. The memory system includes multiple memory modules, each capable of storing data independently. The classification process may involve analyzing image features, such as color, texture, or object recognition, or using metadata tags associated with the images. By grouping images of the same class on the same memory module, the system improves data access efficiency, reduces latency during retrieval, and optimizes storage utilization. This approach is particularly useful in applications requiring fast access to categorized image datasets, such as image search engines, digital libraries, or machine learning training datasets. The invention ensures that related images are physically co-located, enhancing performance for operations like batch processing or parallel retrieval. The system may also include mechanisms for dynamically updating the classification and storage assignments as new images are added or existing classifications are refined.
31. The non-transitory computer-readable medium of claim 27 wherein the search of images includes performing a global descriptor computation, a local descriptor computation, and a verification determination.
The invention relates to image search and retrieval systems, specifically improving the accuracy and efficiency of searching large image databases. The problem addressed is the computational complexity and potential inaccuracy of traditional image search methods, which often rely on either global or local descriptors alone, leading to suboptimal results. The invention enhances image search by combining multiple descriptor computation techniques with a verification step to improve search performance. The system performs a global descriptor computation to capture high-level, broad features of an image, such as overall color distribution or shape. This provides a coarse but fast initial filtering of the image database. Additionally, a local descriptor computation is performed to extract fine-grained details, such as edges, textures, or specific object features, which are critical for distinguishing between similar images. These local descriptors are computed for multiple regions or keypoints within the image to ensure comprehensive coverage. After computing both global and local descriptors, a verification determination is made to refine the search results. This step involves comparing the computed descriptors against stored descriptors in the database to identify the most relevant matches. The verification may include additional checks, such as geometric consistency or descriptor similarity thresholds, to ensure the accuracy of the results. By integrating these three steps—global descriptor computation, local descriptor computation, and verification—the system achieves a more robust and efficient image search process. This approach reduces false positives and improves the precision of image retrieval in large-scale databases.
32. The non-transitory computer-readable medium of claim 27 wherein the comparison includes ranking the search results received from the plurality of memory modules.
This invention relates to a system for improving search efficiency in distributed memory architectures, particularly in environments where multiple memory modules store data that must be searched and ranked. The problem addressed is the inefficiency and complexity of searching across distributed memory modules and combining results, especially when ranking or prioritizing those results is required. The invention provides a method for comparing and ranking search results obtained from multiple memory modules to produce a unified, prioritized output. The system first distributes a search query to a plurality of memory modules, each of which independently processes the query and returns search results. The invention then compares these results, applying a ranking algorithm to determine the most relevant or highest-priority results across all modules. This ranking may be based on factors such as relevance scores, recency, or other predefined criteria. The ranked results are then combined into a single output, allowing users or downstream systems to access the most important information first. The invention ensures that search operations in distributed memory systems are not only efficient but also produce meaningful, prioritized results, improving usability and performance in applications like databases, cloud storage, or distributed computing environments.
33. The non-transitory computer-readable medium of claim 32 , further comprising constructing a final rank of the search results.
A system and method for improving search result ranking in information retrieval processes. The technology addresses the challenge of presenting search results in an order that maximizes relevance and user satisfaction, particularly in scenarios where initial ranking algorithms may not fully account for contextual or dynamic factors. The system processes search results by analyzing user interaction data, such as click-through rates, dwell time, and query reformulation patterns, to refine the ranking of results. It may also incorporate external signals, such as social media engagement or expert-curated content, to enhance relevance. The system constructs a final rank of the search results by applying a machine learning model trained on historical user behavior and content features. This model dynamically adjusts the ranking based on real-time feedback and contextual signals, ensuring that the most relevant results are prioritized. The approach improves search accuracy and user experience by continuously adapting to evolving user preferences and content trends. The system is particularly useful in large-scale search engines, enterprise search applications, and personalized recommendation systems where dynamic ranking adjustments are critical.
Unknown
January 7, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.